Targeting abbreviated medication names with NLP


Mara Alexeev, MD, MPH 1,


Amir Kimia, MD2 Marvin B. Harper, MD2 Assaf Landschaft, M.Sc.3 Al Ozonoff, PhD, CPPS4, 5

1 Department of Pediatrics, Boston Children’s Hospital
2 Division of Emergency Medicine, Boston Children’s Hospital
3 Boston Children’s Hospital
4 Precision Vaccines Program, Boston Children’s Hospital
5 Department of Pediatrics, Harvard Medical School

Introduction

According to The Joint Commission, medication names should not be abbreviated as misinterpretation may lead to administration of incorrect medication. Computerized order entry use eliminates this problem for orders, but clinical notes and narratives are still filled with abbreviations.

Objectives

Identify abbreviated medication names in clinical narratives, using Natural Language Processing, as a first step towards elimination in medical documentation.

Methods

Retrospective chart review of pediatric ED consult notes at a tertiary pediatric center in 2019. We targeted consult notes due to potential differences in expertise between the documenting and reading providers.

Abbreviated, misspelled and true medication names were identified using 2 Natural Language Processing methods:

  1. named-entity recognition (NER) using a pre-trained model called MED7

  2. Regular Expressions (RegEx) used to identify strings likely to be medications given surrounding text context

Results

Selected Abbreviated Medication Names Found
Term Count Potential Meanings
vanc 101 vancomycin
ctx 98 ceftriaxone, Cytoxan
vanco 67 vancomycin
midaz 62 midazolam
ceftaz 39 ceftazidime
lzp 35 lorazepam
amox 32 amoxicillin
norepi 30 norepinephrine
tazo 20 tazobactam
oxc 20 oxcarbazepine, ofloxacin, oxycodone
oxcarb 18 oxcarbazepine
tacro 16 tacrolimus
vgb 15 vigabatrin
ivmp 14 intravenous methylprednisolone
mmf 14 mycophenolate mofetil, maxillomandibular fixation
acei 14 angiotensin converting enzyme inhibitor, acetylcholinesterase inhibitors
phb 13 phenobarbital

Conclusions

Natural Language Processing tools can create libraries of abbreviated medication names used by clinicians. We can then use the output to determine usage frequency and risk of misinterpretation. Abbreviations are not limited to medications; future studies should include more abbreviation types and domain-experts to help interpret domain-specific expressions. Like a spell checker, these libraries could be incorporated into documentation tools in EMR systems to suggest expanded terms.

Funding Sources Alexeev—Biomedical Informatics and Data Science Research Training Program, T15LM007092-30; Ozonoff, Landschaft, Kimia—AHRQ Research Grant 5R01HS026246

Natural Language Processing tools can identify abbreviated medication names.

Libraries of these should be incorporated seamlessly into clinical documentation tools.